{"slug": "medical-ai-doesnt-just-need-bigger-models-it-needs-an-imagenet-for-state", "title": "Medical AI Doesn’t Just Need Bigger Models. It Needs an ImageNet for State Transitions", "summary": "The next frontier for medical AI is building \"world models\" that can predict how a biological state changes in response to an intervention, moving beyond current systems focused on classification or question-answering. To achieve this, the author proposes creating a shared research infrastructure called \"Biomedical TransitionNet,\" analogous to ImageNet, which would standardize and evaluate data on baseline states, interventions, and follow-up states. This infrastructure is considered essential for enabling meaningful progress in understanding and altering biological trajectories, rather than simply developing larger chatbots.", "body_md": "Whoever builds the “state–intervention–transition” dataset for biomedicine may define the next generation of medical AI infrastructure.\n\n**Author:** Jianghui Xiong\n\nMedical AI is moving beyond classification, risk prediction, and question answering.\n\nThe next frontier is not just:\n\n```\nsample → label\n```\n\nor:\n\n```\nquestion → answer\n```\n\nIt is:\n\n```\nstate + action → next state\n```\n\nIn other words:\n\n```\ncurrent biological state + intervention → future biological state\n```\n\nTo build real biomedical world models, we need more than bigger models. We need something analogous to ImageNet — not for images, but for **biological state transitions**.\n\nI will call this idea, for now:\n\n```\nBiomedical TransitionNet\n```\n\nA shared infrastructure for recording, standardizing, and evaluating:\n\n```\nbaseline biological state\n\n- intervention\n- follow-up biological state\n- mechanism evidence\n- uncertainty\n```\n\nThis article explains **why such an infrastructure is needed**, **why it matters**, and **why it is scientifically difficult**.\n\nIt does **not** claim that a complete biomedical world model already exists.\n\n## 1. ImageNet was not just a dataset. It was infrastructure.\n\nWhen people talk about the deep learning revolution in computer vision, they often mention AlexNet, VGG, ResNet, and other neural network architectures.\n\nThat is correct, but incomplete.\n\nOne of the most important enabling factors was **ImageNet**.\n\nImageNet was not merely a large collection of images. Its deeper value was that it gave computer vision a shared coordinate system:\n\n- a common task,\n- a common label hierarchy,\n- common training and test data,\n- common benchmarks,\n- and a way to compare progress across models and institutions.\n\nBefore ImageNet, many computer vision systems were difficult to compare because they were trained and evaluated on different datasets. ImageNet helped the field converge around shared evaluation.\n\nThat is why ImageNet became much more than a database. It became **research infrastructure**.\n\nMedical AI may now need something similar.\n\nBut not an image dataset.\n\nMedicine needs an ImageNet for **state transitions**.\n\n## 2. Medical AI has many models, but not enough transition data\n\nToday, we already have many types of medical AI systems:\n\n- medical large language models,\n- medical question-answering systems,\n- radiology models,\n- pathology models,\n- omics foundation models,\n- virtual cell models,\n- digital twin systems,\n- clinical decision support tools,\n- AI drug discovery platforms.\n\nThese are important.\n\nBut if we think the future of medical AI is only “a bigger medical chatbot”, we may miss the real challenge.\n\nMedicine is not only about answering questions.\n\nMedicine is about understanding and changing biological trajectories.\n\nA clinician does not only ask:\n\n```\nWhat disease does this patient have?\n```\n\nThey also ask:\n\n```\nWhy is this biological state happening?\n\nWhat is driving deterioration?\n\nWhich mechanisms are actionable?\n\nWhich intervention may shift the trajectory?\n\nHow should the response be measured?\n\nWhat if the expected response does not happen?\n\nWhat if an adverse response appears?\n```\n\nThese are not just language problems.\n\nThey are **state transition problems**.\n\nMost medical AI today is still closer to:\n\n```\nsample → label\n```\n\nor:\n\n```\nquestion → answer\n```\n\nBut biomedical world models require something closer to:\n\n```\nstate + action → next state\n```\n\nThat is the key shift.\n\n## 3. What is a biomedical world model?\n\nIn AI, a world model is usually understood as an internal model that helps an agent simulate how the environment changes after an action.\n\nA simple abstraction is:\n\n```\ncurrent state + action → future state\n```\n\nIn robotics, this may mean:\n\n```\nrobot pose + motor command → next scene state\n```\n\nIn autonomous driving, it may mean:\n\n```\ntraffic scene + driving action → future traffic scene\n```\n\nIn biomedicine, the analogous formulation would be:\n\n```\nbiological state + intervention → future biological state\n```\n\nThis could apply at multiple scales:\n\n```\ncell state + perturbation → cellular response\n\ntissue state + treatment → tissue response\n\npatient state + intervention → follow-up state\n```\n\nA biomedical world model should therefore not be understood as a medical chatbot.\n\nIt is not merely:\n\n```\nmedical text in → medical text out\n```\n\nA more meaningful biomedical world model would combine:\n\n```\nstate representation\n\n- intervention representation\n- transition modeling\n- mechanism evidence\n- uncertainty estimation\n- feedback correction\n```\n\nThat is much harder than ordinary medical QA.\n\nAnd it requires a different kind of data.\n\n## 4. Why medicine needs its own ImageNet\n\nIn computer vision, a basic supervised learning unit can often be simplified as:\n\n```\nimage + label\n```\n\nFor biomedical world models, the basic unit should look more like:\n\n```\nbaseline state + action + follow-up state\n```\n\nOr mathematically:\n\n```\nS(t) + A → S(t + Δt)\n```\n\nWhere:\n\n```\nS(t)       = biological state before intervention\n\nA          = action or intervention\n\nS(t + Δt)  = biological state after intervention\n\nΔt         = time interval\n```\n\nThis is fundamentally different from a static medical database.\n\nA biomedical world model does not only need:\n\n- medical images,\n- electronic health records,\n- omics profiles,\n- drug-target databases,\n- clinical notes,\n- literature graphs.\n\nThose are useful, but insufficient.\n\nIt needs structured longitudinal data describing:\n\n```\nwhat the biological state was,\n\nwhat action was taken,\n\nwhat changed afterward,\n\nover what time scale,\n\nwith what evidence,\n\nand with what uncertainty.\n```\n\nThis is why medicine needs something like a **Biomedical TransitionNet**.\n\nNot a direct copy of ImageNet.\n\nA new infrastructure designed for biological state transitions.\n\n## 5. What should one data unit look like?\n\nA conventional supervised learning sample may look like:\n\n```\nx → y\n```\n\nExamples:\n\n```\nimage → diagnosis label\n\nclinical note → ICD code\n\ngenomic variant → risk category\n```\n\nA biomedical world-model sample should look more like:\n\n```\nstate_before\n\n- intervention\n- state_after\n- time_interval\n- evidence_chain\n- uncertainty\n```\n\nA simplified schema might look like this:\n\n```\n{\n\n\"baseline_state\": {\n\n\"molecular\": \"...\",\n\n\"clinical\": \"...\",\n\n\"phenotype\": \"...\",\n\n\"lifestyle\": \"...\",\n\n\"context\": \"...\"\n\n},\n\n\"action\": {\n\n\"type\": \"...\",\n\n\"dose\": \"...\",\n\n\"frequency\": \"...\",\n\n\"duration\": \"...\",\n\n\"mechanism\": \"...\"\n\n},\n\n\"follow_up_state\": {\n\n\"molecular\": \"...\",\n\n\"clinical\": \"...\",\n\n\"phenotype\": \"...\",\n\n\"adverse_events\": \"...\"\n\n},\n\n\"transition\": {\n\n\"direction\": \"...\",\n\n\"magnitude\": \"...\",\n\n\"time_scale\": \"...\",\n\n\"confidence\": \"...\"\n\n},\n\n\"evidence_chain\": {\n\n\"target\": \"...\",\n\n\"pathway\": \"...\",\n\n\"biomarker\": \"...\",\n\n\"phenotype\": \"...\",\n\n\"validation\": \"...\"\n\n}\n\n}\n```\n\nThis is obviously simplified.\n\nBut the principle matters:\n\nA biomedical world model should learn not only:\n\n```\nwhat this sample is\n```\n\nbut:\n\n```\nhow this biological system changed after a defined intervention\n```\n\n## 6. Five layers of a biomedical ImageNet\n\nIf we want to build an ImageNet-like infrastructure for biomedical world models, it should include at least five layers.\n\n## 6.1 State representation\n\nThe first question is:\n\n```\nWhat is the biological state?\n```\n\nA patient state is not just a diagnosis label.\n\nTerms such as:\n\n```\ndiabetes\n\nhypertension\n\naging\n\ninflammation\n\nfatigue\n\nfrailty\n```\n\nare useful, but they are high-level descriptions.\n\nA real biological state may include:\n\n- genome,\n- DNA methylation,\n- transcriptome,\n- proteome,\n- metabolome,\n- immune state,\n- inflammatory state,\n- organ function,\n- microbiome,\n- sleep,\n- activity,\n- diet,\n- medication history,\n- environmental exposure,\n- clinical background.\n\nA simplified representation may be:\n\n```\nindividual_state =\n\nmolecular_state\n\n- pathway_state\n- organ_state\n- phenotype_state\n- lifestyle_context\n- clinical_context\n```\n\nWithout a state representation, a biomedical world model does not know what it is simulating.\n\n## 6.2 Action ontology\n\nA world model needs actions.\n\nIn medicine, actions are complex.\n\nThey may include:\n\n- drugs,\n- supplements,\n- diet,\n- exercise,\n- sleep intervention,\n- stress management,\n- cell therapy,\n- gene therapy,\n- regenerative medicine,\n- combination therapy,\n- N-of-1 personalized intervention.\n\nEven a drug intervention requires many parameters:\n\n```\ndrug name\n\ndose\n\nfrequency\n\nroute\n\nduration\n\ncombination\n\nadherence\n\ncontraindications\n\nadverse events\n```\n\nExercise intervention also requires:\n\n```\ntype\n\nintensity\n\nfrequency\n\nduration\n\nheart-rate zone\n\nrecovery condition\n\nbaseline fitness\n```\n\nIf actions are not standardized, the model cannot learn meaningful transitions.\n\n## 6.3 Transition record\n\nThe core of a biomedical world model is the transition:\n\n```\nbefore → after\n```\n\nExamples:\n\n```\ninflammatory state before intervention → inflammatory state after intervention\n\nDNA methylation age before intervention → DNA methylation age after intervention\n\nmetabolic state before intervention → metabolic state after intervention\n\ntumor state before treatment → tumor state after treatment\n```\n\nWithout follow-up measurement, there is no transition.\n\nWithout transition, there is no world model.\n\nMany medical datasets are still one-time measurements:\n\n```\none-time measurement\n```\n\nBiomedical world models need:\n\n```\nlongitudinal measurement\n```\n\n## 6.4 Evidence chain\n\nA medical model should not only output a probability.\n\nIf a model says:\n\n```\nThis intervention may help.\n```\n\nThat is not enough.\n\nIt should also answer:\n\n```\nWhich targets are involved?\n\nWhich pathways are affected?\n\nWhich abnormal state does this address?\n\nWhich biomarkers can validate the response?\n\nWhich evidence comes from experiments?\n\nWhich evidence comes from clinical data?\n\nWhich part is only model inference?\n\nWhich risks should be monitored?\n```\n\nIn medicine, prediction alone is not sufficient.\n\nA safer output should look more like:\n\n```\nprediction + mechanism + validation + uncertainty\n```\n\nThis is especially important because medical AI should not become an uninspectable black box.\n\n## 6.5 Benchmark task\n\nImageNet helped computer vision because different models could be compared on shared tasks.\n\nBiomedical world models need benchmarks too.\n\nPossible benchmark tasks include:\n\n- cellular perturbation response prediction,\n- gene expression response after drug perturbation,\n- tumor state simulation after treatment,\n- metabolic biomarker response prediction,\n- inflammatory state transition prediction,\n- aging-related biomarker transition prediction,\n- N-of-1 intervention response direction prediction.\n\nBut the metrics cannot be copied directly from image classification.\n\nUseful metrics may include:\n\n```\ndirectional accuracy\n\nmechanistic consistency\n\nbiomarker validation\n\nuncertainty calibration\n\nrisk awareness\n\ncross-context generalization\n```\n\nThis is much harder than top-1 accuracy.\n\nBut medicine requires it.\n\n## 7. Related progress: promising, but still early\n\nTo be scientifically careful, we should not pretend that complete biomedical world models already exist.\n\nThey do not.\n\nBut several related directions are emerging.\n\n## 7.1 ImageNet as an infrastructure analogy\n\nImageNet and ILSVRC showed how large-scale, standardized datasets and benchmarks can accelerate a field.\n\nHowever, ImageNet is a benchmark for image classification and detection.\n\nIt is not equivalent to what biomedicine needs.\n\nHere, ImageNet is used only as an infrastructure analogy.\n\nThe biomedical version must be longitudinal, dynamic, intervention-aware, and mechanism-sensitive.\n\n## 7.2 World Models in AI\n\nHa and Schmidhuber’s **World Models** is a representative work in AI world modeling.\n\nIts key idea is that an agent can learn an internal model of the environment and use it to simulate future states.\n\nMedicine cannot directly copy this setting.\n\nA human body is not a game environment.\n\nClinical intervention cannot be freely explored by trial and error.\n\nBut the abstraction:\n\n```\nstate + action → future state\n```\n\nis still useful for thinking about medical AI.\n\n## 7.3 Virtual cells and perturbation response\n\nArc Institute’s **State** model is a recent example of virtual-cell modeling.\n\nIt aims to predict how cells respond to drugs, cytokines, or genetic perturbations. Public descriptions indicate that State was trained on large-scale observational and perturbational single-cell data.\n\nThis is important because it directly touches the pattern:\n\n```\ncell state + perturbation → cellular response\n```\n\nHowever, State is primarily a cellular-level model.\n\nIt should not be confused with a complete patient-level biomedical world model.\n\n## 7.4 Medical World Model for tumor evolution\n\nRecent work using the term **Medical World Model**, such as MeWM, explores generative simulation of tumor evolution under treatment conditions.\n\nThis is relevant because it moves medical AI from static recognition toward treatment-conditioned disease dynamics.\n\nBut this direction is still early.\n\nIt should not be interpreted as a general solution to biomedical world modeling.\n\n## 7.5 Digital twins and virtual physiological systems\n\nLong before today’s AI world-model terminology, fields such as computational physiology, systems biology, virtual physiological systems, and digital twins already tried to connect biological structure, mechanism, dynamics, and measurable outputs.\n\nThat tradition matters.\n\nA good biomedical world model should not be just a black-box predictor.\n\nIt should connect:\n\n```\nstate\n\nmechanism\n\ndynamic change\n\nmeasurement\n\nfeedback\n```\n\nToday’s biomedical world models can be seen as an extension of this older systems-modeling tradition into the era of AI, multi-omics, real-world data, and large-scale computation.\n\n## 8. Why steerability matters\n\nA biomedical world model that only predicts is not enough.\n\nA model may predict that a patient’s risk is increasing.\n\nBut medicine needs more than that.\n\nIt needs to ask:\n\n```\nWhich state can be measured?\n\nWhich abnormality can be explained?\n\nWhich intervention can be described?\n\nWhich transition can be tested?\n\nWhich deviation can be traced?\n\nWhich failure can be corrected?\n```\n\nThis is why I emphasize **steerability**.\n\nGoing forward, I will use the name:\n\n```\nSteeraMed: A Steerable Biomedical World Model\n```\n\nWebsite:\n\n```\nhttps://SteeraMed.com\n```\n\nThe earlier preprint name was:\n\n```\nSEWO / Steerable Medicine World Model\n```\n\nor in Chinese:\n\n```\n可驾驭医学世界模型\n```\n\nWhenever I mention SEWO / 可驾驭医学世界模型, it should be understood together with the new unified naming:\n\n```\nSteeraMed: A Steerable Biomedical World Model\n```\n\nThe idea behind SEWO / SteeraMed is that biomedical world models should not only pursue predictive accuracy. They should also support:\n\n- state definition,\n- intervention description,\n- transition hypothesis,\n- mechanism audit,\n- deviation tracing,\n- uncertainty inspection,\n- expert steering,\n- and iterative correction.\n\nThe related ideas were introduced in the preprint:\n\n```\nWorld Models for Biomedicine: A Steerability Framework\n```\n\nand are also presented at:\n\n```\nhttps://steerable.world\n```\n\nImportant clarification:\n\nSEWO / SteeraMed is **not** a clinically validated treatment system.\n\nIt is **not** a medical device.\n\nIt is better understood as a structural framework and evidence-chain design principle for future biomedical world models.\n\nThe key question is not only:\n\n```\nCan the model predict?\n```\n\nbut:\n\n```\nCan researchers and clinicians inspect, question, correct, and steer the model within clearly defined boundaries?\n```\n\n## 9. Why longevity medicine may be one entry point\n\nBiomedical world models could start from many areas:\n\n- oncology,\n- cardiovascular disease,\n- metabolic disease,\n- immunology,\n- neurodegeneration,\n- drug discovery,\n- virtual cells,\n- longevity medicine.\n\nLongevity medicine is not the only entry point.\n\nBut it is an interesting one.\n\nWhy?\n\n## 9.1 Aging is a continuous state\n\nAging is not a single disease label.\n\nIt is a continuous, multi-system biological process involving:\n\n- inflammation,\n- metabolism,\n- immunity,\n- epigenetics,\n- mitochondrial function,\n- proteostasis,\n- stem-cell exhaustion,\n- cellular senescence,\n- organ function decline.\n\nThat makes it naturally suitable for state modeling.\n\n## 9.2 Longevity medicine requires repeated measurement\n\nLongevity medicine is not a one-time diagnostic event.\n\nIt depends on repeated measurement over time.\n\nA useful intervention must be evaluated through:\n\n```\nbaseline state → intervention → follow-up state\n```\n\nThis is exactly the structure needed for biomedical world modeling.\n\n## 9.3 Interventions are diverse\n\nLongevity-related interventions may include:\n\n- diet,\n- exercise,\n- sleep,\n- supplements,\n- drugs,\n- cell therapy,\n- regenerative medicine,\n- stress management,\n- environmental exposure management.\n\nThis provides a rich action space.\n\n## 9.4 Individual responses vary\n\nThe same intervention may produce different responses in different people.\n\nThat means longevity medicine cannot rely only on average effects.\n\nIt needs N-of-1 style transition data:\n\n```\nindividual state → intervention → individual transition\n```\n\nEach well-structured N-of-1 intervention can be seen as a small world-model experiment.\n\n## 10. Engineering implications\n\nFrom an engineering perspective, the biomedical ImageNet is not just a dataset.\n\nIt is a data infrastructure problem.\n\nIt requires:\n\n- data collection,\n- data standardization,\n- multimodal integration,\n- time-series modeling,\n- intervention encoding,\n- causal confounding control,\n- privacy protection,\n- benchmark design,\n- safety boundaries,\n- evidence-chain tracking.\n\nA simplified loop may look like:\n\n```\nmeasure state\n\n↓\n\nstandardize state representation\n\n↓\n\nrecord intervention\n\n↓\n\nmeasure follow-up state\n\n↓\n\nconstruct transition sample\n\n↓\n\ntrain / evaluate world model\n\n↓\n\ngenerate testable hypothesis\n\n↓\n\nrepeat and correct\n```\n\nThis is not a static dataset.\n\nIt is a data flywheel.\n\n## 11. Main challenges\n\nThis is scientifically and technically difficult.\n\nSome of the main challenges include:\n\n## 11.1 Biological state is complex\n\nA human state cannot be compressed into one label.\n\nWe need ways to represent multi-omics, clinical metrics, imaging, lifestyle, symptoms, environmental exposure, and medical history as computable state variables.\n\n## 11.2 Interventions are hard to standardize\n\nDrugs, exercise, diet, sleep, supplements, and cell therapies all have complex parameters.\n\nWithout action standardization, transition learning will be noisy.\n\n## 11.3 Follow-up data is scarce\n\nMost medical data is not collected as structured pre/post intervention transition data.\n\nThis requires new data collection workflows.\n\n## 11.4 Causal confounding is serious\n\nIn the real world, people often change many things at once:\n\n```\ndiet\n\nexercise\n\nsleep\n\nmedication\n\nsupplements\n\nstress\n```\n\nAttributing a state change to one factor is difficult.\n\nThis requires careful study design and statistical methods.\n\n## 11.5 Safety and ethics are central\n\nA biomedical world model cannot freely experiment like a game-playing agent.\n\nAny intervention-related model must clearly distinguish:\n\n```\nresearch hypothesis\n\nhealth-management suggestion\n\nclinical decision support\n\nmedical recommendation\n\nvalidated therapy\n```\n\nClinical use would require prospective validation, safety evaluation, ethical review, regulatory review where applicable, and professional oversight.\n\n## 11.6 Open standards and business incentives may conflict\n\nIf everything is closed, the field cannot build shared benchmarks.\n\nIf everything is open, companies may lack incentives to invest.\n\nA practical ecosystem will need a balance among:\n\n```\nopen benchmarks\n\nprivacy protection\n\ncommercial incentives\n\nscientific collaboration\n```\n\n## 12. A minimal viable direction\n\nA biomedical ImageNet should not begin by trying to simulate the entire human body.\n\nA more realistic path is to start with minimal viable tasks.\n\nExamples:\n\n- cellular perturbation response prediction,\n- tumor state change after treatment,\n- metabolic biomarker response prediction,\n- inflammatory state transition prediction,\n- DNA methylation age transition,\n- N-of-1 longevity intervention tracking.\n\nA minimal task should define:\n\n```\n1. state variables\n2. intervention variables\n3. follow-up time\n4. transition metrics\n5. benchmark task\n6. safety boundary\n```\n\nStart narrow.\n\nMake it measurable.\n\nMake it repeatable.\n\nMake it auditable.\n\nThen scale.\n\n## 13. Whoever defines state, action, and transition may define the field\n\nMedical AI will still need better models.\n\nBut bigger models alone cannot solve the problem of biomedical state transition learning.\n\nThe scarce asset is the infrastructure that allows models to learn:\n\n```\nhow life systems change after intervention\n```\n\nFuture platform-level medical AI companies may not be the ones with the largest language models.\n\nThey may be the ones that can build the strongest data flywheel:\n\n```\nmeasure biological state\n\nstandardize interventions\n\nrecord follow-up changes\n\nconstruct mechanism evidence chains\n\nevaluate transition models\n\nrepeat\n```\n\nWhoever defines `state`\n\ndefines what medical AI can see.\n\nWhoever defines `action`\n\ndefines how medical AI understands intervention.\n\nWhoever defines `transition`\n\ndefines how medical AI learns biological change.\n\nWhoever defines the benchmark defines how the field measures progress.\n\n## Conclusion\n\nImageNet helped machines learn to see the world.\n\nA biomedical ImageNet should help AI learn how life responds to intervention.\n\nThat does not mean replacing clinicians.\n\nIt means building a scientific infrastructure where models can learn:\n\n```\nhow states form\n\nhow interventions act\n\nhow systems transition\n\nhow evidence is validated\n```\n\nThe next decade of medical AI may not be limited by model size alone.\n\nIt may be limited by the lack of a shared infrastructure for biological state transitions.\n\nThat is the real opportunity.\n\n## References\n\nDeng J, Dong W, Socher R, Li LJ, Li K, Fei-Fei L.\n\n**ImageNet: A Large-Scale Hierarchical Image Database.** CVPR. 2009.\n\n[https://ieeexplore.ieee.org/document/5206848](https://ieeexplore.ieee.org/document/5206848)Russakovsky O, Deng J, Su H, et al.\n\n**ImageNet Large Scale Visual Recognition Challenge.** International Journal of Computer Vision. 2015.\n\n[https://arxiv.org/abs/1409.0575](https://arxiv.org/abs/1409.0575)ImageNet official website.\n\n[https://www.image-net.org/](https://www.image-net.org/)Ha D, Schmidhuber J.\n\n**World Models.** 2018.\n\n[https://worldmodels.github.io/](https://worldmodels.github.io/)Arc Institute.\n\n**Arc Institute’s first virtual cell model: State.**\n\n[https://arcinstitute.org/news/virtual-cell-model-state](https://arcinstitute.org/news/virtual-cell-model-state)**Predicting cellular responses to perturbation across diverse contexts with State.** bioRxiv. 2025.\n\n[https://www.biorxiv.org/content/10.1101/2025.06.26.661135v1](https://www.biorxiv.org/content/10.1101/2025.06.26.661135v1)Yang Y, Wang ZY, Liu Q, et al.\n\n**Medical World Model: Generative Simulation of Tumor Evolution for Treatment Planning.** arXiv.\n\n[https://arxiv.org/abs/2506.02327](https://arxiv.org/abs/2506.02327)IEEE Transactions on Biomedical Engineering.\n\n**Digital Twins / AI World Models.**\n\n[https://www.embs.org/tbme/research-highlights/digital-twins-ai-world-models/](https://www.embs.org/tbme/research-highlights/digital-twins-ai-world-models/)Acosta JN, Falcone GJ, Rajpurkar P, Topol EJ.\n\n**Multimodal biomedical AI.** Nature Medicine. 2022.\n\n[https://www.nature.com/articles/s41591-022-01981-2](https://www.nature.com/articles/s41591-022-01981-2)Xiong J.\n\n**World Models for Biomedicine: A Steerability Framework.** Preprints.org. 2026.\n\n[https://www.preprints.org/manuscript/202605.0366](https://www.preprints.org/manuscript/202605.0366)\n\nDOI:[https://doi.org/10.20944/preprints202605.0366.v1](https://doi.org/10.20944/preprints202605.0366.v1)SteeraMed: A Steerable Biomedical World Model.\n\n[https://steerable.world](https://steerable.world)\n\n## Disclaimer\n\nThis article is for research, technical, and industry discussion only.\n\nIt is not medical advice, diagnostic advice, or treatment advice.\n\nAny biomedical world model intended for clinical use would require prospective validation, safety evaluation, ethical review, regulatory review where applicable, and professional clinical oversight.\n\n", "url": "https://wpnews.pro/news/medical-ai-doesnt-just-need-bigger-models-it-needs-an-imagenet-for-state", "canonical_source": "https://dev.to/jxiong/medical-ai-doesnt-just-need-bigger-models-it-needs-an-imagenet-for-state-transitions-28n3", "published_at": "2026-05-19 00:15:48+00:00", "updated_at": "2026-05-19 00:30:52.754978+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "research", "data", "science"], "entities": ["Jianghui Xiong", "ImageNet", "AlexNet", "VGG", "ResNet"], "alternates": {"html": "https://wpnews.pro/news/medical-ai-doesnt-just-need-bigger-models-it-needs-an-imagenet-for-state", "markdown": "https://wpnews.pro/news/medical-ai-doesnt-just-need-bigger-models-it-needs-an-imagenet-for-state.md", "text": "https://wpnews.pro/news/medical-ai-doesnt-just-need-bigger-models-it-needs-an-imagenet-for-state.txt", "jsonld": "https://wpnews.pro/news/medical-ai-doesnt-just-need-bigger-models-it-needs-an-imagenet-for-state.jsonld"}}